Recently, there is a surge of recommender system to alleviate the Internet information overload. A number of recommendation techniques have been proposed for many applications, among which music recommendation is a kind of popular Internet services. Unlike other recommendation services, music recommendation needs to consider the interaction and content information, as well as the inherent correlation and feedback among music playlist. Thus, in this paper, we model music recommendation as a Markov Decision Process, and consider the music recommendation as a playlist recommendation task. Along this line, we propose a novel reinforcement learning based model, called RLWRec, to exploit the optimal strategy of playlist. Two novel strategies are designed to solve the curse of state space and efficient music recommendation. Experiments on real dataset validate the effectiveness of our proposed method.